Member of Technical Staff
The Token Company
Member of Technical Staff
The Token Company
AI middleware for semantic prompt compression
The Token Company (YC W26, HF0 S26) · San Francisco, CA
The Token Company is a seed stage startup in San Francisco. We have raised $12M from First Round Capital, NEA, Y Combinator and SV Angel, alongside the founders of Dropbox, Slack, Supercell and Hugging Face, and key people from OpenAI, xAI and DoorDash.
We train machine learning models that compress raw LLM inputs before they reach the expensive model. Our models learn which parts of an input are redundant and strip them out, cutting inference costs for the scale-ups and enterprises building on LLMs. We are a team of five with a tight research and product focus, and we ship what we train.
As a Member of Technical Staff on the research team, you own the model training stack end to end: data, architecture, training, evaluation, and shipping compression models into production. This is a from-scratch model training role. You will spend your time designing and training models on NVIDIA B200s, not wiring up RAG pipelines or prompting other people's APIs.
You have trained models from scratch. You have owned data, architecture and a training loop yourself, ideally in a research lab, a startup or a scale-up. Not only fine-tuning or calling an API.
You are strong in ML fundamentals. You are fluent with transformers and comfortable turning a paper or a rough idea into a real training run.
You learn fast. You get to the core of a hard problem quickly and iterate on new ideas without hand-holding.
You care about production. You would rather see your model cut real inference costs for real users than add another line to a publication list.
Pretrained a transformer, or done serious post-training or RL on one
Built a novel architecture or training method with results to back it up
Shipped a model you trained into something people actually use
Your experience is mostly RAG, agents, prompt engineering, or fine-tuning existing models through an API
You want to publish papers more than you want to ship models
Competitive base salary plus significant equity. We also cover housing, food, laundry and cleaning in San Francisco, sponsor visas, and give you the resources and a real seat to build the research team around you.
thetokencompany.com
AI middleware for semantic prompt compression
Token Company (Otsofy) offers an AI middleware API that semantically compresses LLM prompts to reduce token usage, latency, and costs. It uses a fast model called bear-1 to preprocess inputs before sending them to LLMs like GPT or Claude, preserving meaning while removing nonessential tokens. The API is a drop-in integration for production use and aims to be a neutral efficiency layer across models, helping teams run larger prompts within fixed compute budgets. Its goal is to become a core part of the AI stack by improving efficiency for every LLM request.
Explore more opportunities at The Token Company